MallaNet residual branch merge convolutional neural network with homogeneous filter capsules for Devanagari character recognition

The Devanagari script's complex character set and handwriting variability pose significant challenges for handwritten character recognition (HCR). This study aims to develop a robust deep learning model, MallaNet, to achieve high accuracy in recognizing Devanagari characters by leveraging multiscale feature extraction and preserving spatial hierarchies. We introduce the Residual Enhanced Branching and Merging Convolutional Neural Network with Homogeneous Filter Capsules (MallaNet), an optimized deep learning model designed to address these complexities. Extending the Branching and Merging Convolutional Network with Homogeneous Vector Capsules (BMCNNwHVCs), our model integrates optimized residual blocks, refined Homogeneous Filter Capsule (HFC) layers, and a merging layer to capture multiscale features and preserve spatial hierarchies, critical for distinguishing visually similar characters. Trained in the Devanagari Handwritten Character Dataset (DHCD), comprising 92,000 images across 46 classes, MallaNet achieves a test accuracy of 99.71%, macro-average F1-score of 99.71%, closely approaching the highest reported accuracy of 99.72% while utilizing 56.41% fewer parameters (17M vs. 39M) and surpassing previous benchmarks of 98.47% and 99.16%, enhancing optical character recognition (OCR) for regional scripts and supporting document digitization and cultural preservation with improved efficiency.